Deep learning-based electric dust removal variable frequency power supply working condition regulation method

Abstract

The present application relates to the technical field of electric dust removal control, and discloses an electric dust removal variable frequency power working condition regulation and control method based on deep learning, which comprises the following steps: based on macroscopic operation state, an evaluation network forms a delay sampling point number, an embedding dimension and a dynamic evolution threshold; original high-frequency voltage and current signals collected synchronously are processed to form microscopic electrical characteristics and calculate a finite-time maximum Lyapunov index; a gating coefficient is determined according to the dynamic evolution threshold, microscopic electrical characteristics are fused with macroscopic state to splice a cross-scale feature vector; the cross-scale feature vector, a feedback control state value and a current power supply mode are input into a multi-task feedforward neural network to output a microscopic working condition classification result and a control parameter initial setting value, and a control instruction is formed through deterministic boundary constraint verification and coding. Through adaptive fusion of multi-source features and multi-task reasoning, and in combination with software and hardware boundary constraints, the present application improves the system regulation and control accuracy and safety.

Description

Technical Field

[0001] This invention relates to the field of electrostatic precipitator control technology, specifically a method for regulating the operating conditions of electrostatic precipitator variable frequency power supplies based on deep learning. Background Technology

[0002] Electrostatic precipitator variable frequency power supplies are affected by both macroscopic conditions such as boiler load and flue gas temperature, and microscopic electrical characteristics such as high-frequency voltage and current during operation. Existing control methods struggle to adaptively extract phase space reconstruction parameters when processing multi-source data, resulting in an inability to effectively integrate macroscopic operating conditions with microscopic electrical characteristics, leading to biases in the data foundation for equipment condition assessment.

[0003] Due to insufficient fusion of cross-scale features, the control system struggles to accurately output micro-condition classification results when facing complex power supply conditions such as spark precursors and back corona. Current control strategies often rely on fixed judgment logic or single-variable feedback, failing to synchronously generate control parameter setpoints and adaptively adjust the power supply mode in the control loop. This results in a lag in the adjustment of the variable frequency power supply output parameters, affecting the operational stability of the dust removal equipment.

[0004] When introducing deep learning algorithms for operating condition judgment and control parameter setting, the inherent uninterpretability of the algorithms poses potential safety risks to the system. Existing control mechanisms lack deterministic boundary constraints on output power when algorithmic reasoning deviates or classification confidence is insufficient. Furthermore, when facing sudden situations such as secondary overcurrent, they lack asynchronous hardware protection logic independent of the software algorithm, which can easily lead to delays in blocking actions due to algorithmic intervention, resulting in insufficient overall safety of the power control system. Summary of the Invention

[0005] The technical problem solved by this invention is that existing electrostatic precipitator frequency converters are unable to accurately identify the coupling relationship between microscopic electrical characteristics and macroscopic operating states, resulting in deviations in operating condition classification and control parameter settings, and insufficient system response speed and control safety under abnormal operating conditions.

[0006] To address the above problems, the present invention provides the following technical solution: This invention provides a deep learning-based method for controlling the operating conditions of an electrostatic precipitator variable frequency power supply, comprising: The evaluation network is based on macroscopic operational state data to determine the number of delayed sampling points, embedding dimension, and dynamic evolution threshold; The original high-frequency secondary voltage sampling signal and secondary current sampling signal acquired synchronously are processed to form a one-dimensional net voltage timing signal and secondary current correlation features; The phase space of the one-dimensional net voltage time-series signal is reconstructed by the number of delayed sampling points and the embedding dimension, forming a finite-time maximum Lyapunov exponential input feature, which, together with the secondary current-related feature, constitutes the microscopic electrical feature. The micro-feature gating coefficients are determined based on the input features and dynamic evolution thresholds, and the micro-electrical features and macro-operational status data are fused to form a cross-scale feature vector; The cross-scale feature vector, feedback control state value, and current power supply mode are input into the control network to obtain the micro-operating condition classification results and the initial set values ​​of the control parameters. The initial set values ​​of the control parameters are verified by deterministic boundary constraints and encoded to form control parameter commands.

[0007] By adopting the above technical solution, the phase space reconstruction parameters and evolution thresholds are dynamically generated by the evaluation network. The one-dimensional net voltage time series signal is mapped to a high-dimensional phase space to extract the maximum Lyapunov exponent in a finite time. Combined with the secondary current correlation characteristics, the nonlinear dynamic evolution trend of the micro-operating condition is reflected. The fusion weight of micro- and macro-features is adjusted by the micro-feature gating coefficient, so that the parameter setpoint of the control network output takes into account both the global operating condition changes and the partial discharge characteristics. Combined with deterministic boundary constraint verification, the control command is prevented from going out of bounds, thus realizing the data-driven control and safe operation limit of the electrostatic precipitator power supply.

[0008] Furthermore, the macroscopic operating status data includes boiler load, flue gas temperature at the dust collector inlet, and flue gas flow rate, and the evaluation network is a multilayer sensor. The candidate delayed sampling points, candidate embedding dimension, and candidate dynamic evolution threshold formed by the same macroscopic operating state update constitute an indivisible candidate dynamic parameter set; The stability condition is met when the number of candidate delayed sampling points and the candidate embedding dimension are the same in two consecutive macroscopic operating state update cycles, and the difference in the candidate dynamic evolution threshold does not exceed the preset threshold change tolerance. Once the candidate dynamic parameter set meets the stability condition, the preset historical data length requirement, and the preset control update cycle boundary condition, it will be activated as a whole in the next control update cycle. The three parameters will be used as the current effective number of delayed sampling points, the current effective embedding dimension, and the current effective dynamic evolution threshold, respectively.

[0009] By adopting the above technical solution, macroscopic physical quantities are used as inputs to the multilayer perceptron to characterize the system boundary conditions. Stability conditions and threshold change tolerances are set to filter disturbances in the state update process. This ensures that the dynamic parameter set takes effect as a whole after meeting the boundary conditions of historical data length and update cycle. It also ensures that the number of sampling points and embedding dimension required for feature extraction remain matched and stable in the time series, avoiding phase space reconstruction divergence caused by parameter mutations, which could lead to abnormal fluctuations in the output of the subsequent control network.

[0010] Furthermore, each consecutive effective historical switching cycle is pre-divided, and the periodic average voltage of each switching cycle is calculated. The phase residual is obtained by subtracting the corresponding periodic average voltage from each phase sample value. The arithmetic mean of the phase residuals of the same phase sample point number is calculated to form a phase ripple template. The net voltage sample value is formed by subtracting the phase ripple template value of the corresponding phase from the current original high-frequency secondary voltage sample value. The net voltage sample value is then formed by digital low-pass filtering and preset integer multiple downsampling to form a one-dimensional net voltage timing signal.

[0011] By adopting the above technical solution, the phase residual is calculated based on the deviation between the average voltage and the phase sample value within the historical switching cycle. A phase ripple template is established by using multi-cycle arithmetic averaging to suppress random noise. The template value is subtracted from the real-time high-frequency secondary voltage sample value to remove the periodic fluctuations caused by the switching frequency. After filtering and downsampling, a one-dimensional net voltage time sequence signal that can reflect the nonlinear characteristics of the electric field inside the dust collector is obtained, which improves the purity of the input data for subsequent phase space reconstruction.

[0012] Furthermore, after the secondary current sampling signal is digitally low-pass filtered and downsampled by the same factor, it is aligned with the one-dimensional net voltage timing signal according to the sampling sequence number or timestamp. A time sliding window corresponding to the phase space reconstruction is established. The mean value of the secondary current is calculated within the time sliding window. The slope of the first-order linear fitting of the secondary current sample value and the sampling time is taken as the rate of change of the secondary current. The zero-delay Pearson linear correlation coefficient between the one-dimensional net voltage time sequence signal and the aligned secondary current sample signal is taken as the voltage-current correlation feature. The three together constitute the secondary current correlation feature.

[0013] By adopting the above technical solution, a time-aligned sliding calculation window is established to ensure that multi-source electrical signals are evaluated under the same dynamic state. The rate of change of secondary current is quantified by first-order linear fitting of the skew force. The synchronous correlation between the fluctuations of net voltage and secondary current is evaluated by zero-delay Pearson linear correlation coefficient. The mean and the trend of change are combined to form multi-dimensional secondary current correlation features, which enhances the algorithm's ability to perceive and identify early states of sparks and back corona.

[0014] Furthermore, a historical voltage data buffer queue for storing the net voltage signal is constructed, and a target net voltage sampling point is selected within the same continuous valid data segment of the historical voltage data buffer queue. Taking the target net voltage sampling point as the end, select historical sampling points with the current effective embedding dimension reduced by one in the historical direction according to the current effective number of delayed sampling points, and form a multi-dimensional phase space state vector. The effective state vectors are arranged in chronological order to form a multidimensional phase space trajectory matrix.

[0015] By adopting the above technical solution, a historical voltage data buffer queue is established to maintain the temporal integrity of the net voltage signal. The sampling points are limited to continuous effective data segments to prevent data discontinuity from introducing human interference features. Based on the current number of effective delayed sampling points and the embedding dimension, a one-dimensional time series signal is extracted to construct a phase space state vector, which is then arranged to generate a multi-dimensional phase space trajectory matrix. This transforms the univariate time series into a topological structure containing the system's evolution trajectory, providing a data foundation for the extraction of dynamic features.

[0016] Furthermore, the maximum Lyapunov exponent in finite time is calculated from the multidimensional phase space trajectory matrix. Based on the validity status of the calculation result, the input feature is formed by processing the result with valid values, the most recent valid value, or risk substitution values. When processing risk substitute values, the micro-feature gating coefficient is set to the preset maximum gating value; When risk substitution value processing is not performed, the same standardization processing is performed on the input features and the current effective dynamic evolution threshold, and the micro-feature gating coefficient is determined by the Sigmoid function based on the difference between the two. After the micro-feature gating coefficient is determined, the input features, secondary current related features and macro-operating state data are respectively standardized. The standardized micro-electrical features are multiplied by the micro-feature gating coefficient, and the standardized macro-operating state data are multiplied by one and subtracted from the micro-feature gating coefficient. The results are then concatenated to form a seven-dimensional cross-scale feature vector.

[0017] By adopting the above technical solution, the maximum Lyapunov exponent in finite time is calculated using the multidimensional phase space trajectory matrix to quantify the divergence rate of the electrical state. A risk substitution mechanism is set to deal with data failure caused by calculation anomalies. The difference between the input features and the dynamic evolution threshold is smoothly transformed by the Sigmoid function to form the micro-feature gating coefficient. The weight allocation of micro-electrical features is adaptively improved when the features fluctuate. The dimensional differences of cross-scale data are eliminated through standardization processing, and the fusion feature vector is spliced ​​to form a fusion feature vector that is conducive to the stable inference of the feedforward neural network.

[0018] Furthermore, the control network is a multi-task feedforward neural network that includes a shared input layer, a shared hidden layer, a classification branch, and a regression branch; The inputs to the control network include a seven-dimensional cross-scale feature vector, three feedback control state values ​​corresponding to the fundamental voltage amplitude, pulse power supply frequency, and pulse duty cycle, respectively, and a three-dimensional one-hot encoding of the current power supply mode. The classification branch outputs the classification probabilities of normal corona, spark precursor, and reverse corona, which are then verified by confidence to form the micro-operating condition classification result. The initial settings for the fundamental voltage amplitude, pulse power supply frequency, and pulse duty cycle of the regression branch output.

[0019] By adopting the above technical solution, a multi-task architecture is used to process classification and regression tasks simultaneously. The shared layer is responsible for extracting the general pattern from the cross-scale feature vector. The classification branch outputs the probability of normal corona and abnormal states and confirms the working condition status with confidence verification. The regression branch directly outputs the control parameter setting value corresponding to the current working condition. The feedback control state value and the current power supply mode code are input into the network to realize closed loop and state perception, thereby improving the response efficiency of joint control of multi-objective parameters under complex working conditions.

[0020] Furthermore, the average power supply is determined by the power supply mapping table; When there are signs of sparking, it is forbidden to increase the fundamental voltage amplitude and average power supply. It is permissible to reduce the fundamental voltage amplitude, reduce the pulse duty cycle, or enter a low-power pulse operation state in pulse power supply mode. If the current power supply mode is not pulse power supply mode and the average power supply power of the preset low-power pulse control parameter group is not higher than the current average power supply power, the pulse power supply mode and the preset low-power pulse control parameter group are respectively regarded as candidate power supply mode and target control parameter. During back corona discharge, reduce the average power supply or pulse duty cycle; If the current power supply mode is not intermittent and the average power supply power of the preset safety control parameter corresponding to the intermittent power supply mode is lower than the current average power supply power, the intermittent power supply mode and the preset safety control parameter shall be used as the candidate power supply mode and the target control parameter, respectively. When the power supply mode switching is triggered by the spark precursor state or the back corona state, the micro-condition classification result continuously reaches the preset mode switching cycle number, and the current control system is not in the output blocking state. The preset equipment safety conditions corresponding to the candidate power supply mode are met, and the candidate power supply mode and the target control parameters meet the preset theoretical continuous effective power supply duration requirements.

[0021] By adopting the above technical solutions, power reduction control rules are established to suppress discharge for spark precursors and back corona. When conventional adjustment cannot meet the requirements, low-power pulses or intermittent modes are generated as candidate modes. The operating condition classification results are required to be continuously triggered in multiple cycles to filter out misjudgment interference. Output blocking status checks and equipment safety and power supply duration boundary verification are enforced to ensure that the power supply mode switching decision can curb the development of electrical faults and maintain the hardware safety of the system's basic operation.

[0022] Furthermore, deterministic boundary constraint verification includes finite value judgment, physical quantity convertibility judgment, single maximum change limit, absolute value limit, mode switching condition verification, and abnormal state interception. In the abnormal state interception, when the system detects the existence of any of the following trigger states: The following trigger states are triggered: invalid secondary current related characteristics, invalid maximum Lyapunov exponent in finite time, insufficient classification confidence, inconsistent control action direction, external macroscopic input distribution, spark precursor classification result, control network inference timeout, abnormal control parameters, deterministic boundary constraint verification timeout, and hardware blockade. Increasing the fundamental voltage amplitude is prohibited. In pulse power supply mode or intermittent power supply mode, the average power supply corresponding to the control parameter combination should not exceed the average power supply corresponding to the actual control parameters used in the previous control update cycle.

[0023] By adopting the above technical solutions, a deterministic constraint system containing numerical boundaries and physical rules is established at the output end of the neural network to compensate for the uncontrollable risks of the model. In the abnormal state interception stage, the fault triggering states of multiple dimensions such as feature calculation and network inference timeout are monitored. If any triggering state is detected to be valid, the instruction channel to increase the output power is cut off. In special power supply mode, the power not to increase rule is executed, thus building a deterministic safety control defense line for system operation.

[0024] Furthermore, the method is applied to a hardware system including an inverter bridge, a fast comparator, and a hardware latch circuit, and the method also includes hardware asynchronous protection: When the secondary current exceeds the hardware overcurrent protection threshold, the fast comparator directly outputs a hardware blocking signal to asynchronously cut off the gate drive of the inverter bridge switching device, and the hardware latch circuit maintains the hardware latch blocking state. Hardware asynchronous protection always runs independently and in parallel. Its protection priority is higher than control network inference, deterministic boundary constraint verification and software control parameter instructions. Deep learning algorithms cannot shield, delay or cover hardware blocking signals.

[0025] By adopting the above technical solutions, a hardware protection circuit independent of the deep learning computing platform is set up to deal with extreme overcurrent faults. A fast comparator is used to directly cut off the drive signal of the switching device to achieve microsecond-level hardware protection. A hardware latching circuit is used to maintain the blocked state to prevent repeated false triggering caused by operating condition oscillations. The hardware protection logic is set to have the highest priority and is not affected or covered by software algorithm instructions, thus ensuring the safety of the core components of the main circuit of the frequency converter power supply from the physical level.

[0026] Furthermore, after the hardware latch block state is valid, the hardware block trigger state is recorded, the unexecuted power supply mode switching request is revoked, the use of the control network to generate new control parameter instructions is prohibited, the phase ripple template used to process the formation of the one-dimensional net voltage timing signal is marked as invalid, and a data continuity interruption flag is written into the historical data sequence used for phase space reconstruction. After the hardware latch is released, the phase ripple template is reconstructed according to the preset safety control parameters or mode switching safety control parameters corresponding to the current power supply mode, and continuous and valid historical data for phase space reconstruction is accumulated. Before the preset template reconstruction and recovery conditions are met, the control network is not used to generate new control parameter instructions.

[0027] By adopting the above technical solution, during the hardware latching block, the failed intermediate state is cleared and the lingering instructions are revoked. The phase ripple template for processing the net voltage signal is invalidated and an interrupt flag is written into the historical data sequence. This avoids invalid data containing open-circuit noise interfering with the phase space reconstruction stage. After the hardware block is lifted, template reconstruction and effective historical data accumulation are forcibly executed. Transitional power supply is performed using preset safety parameters until the system regains continuous timing characteristics before the control network takes over. This achieves a safe logical connection between hardware protection and software control.

[0028] This invention provides a method for controlling the operating conditions of an electrostatic precipitator variable frequency power supply based on deep learning. It has the following beneficial effects: 1. This invention evaluates the number of delayed sampling points and the embedding dimension generated by the network based on macroscopic operating state data. Based on this, it reconstructs the one-dimensional net voltage timing signal and secondary current characteristics to form microscopic electrical features. The microscopic feature gating coefficient is determined according to the dynamic evolution threshold for data fusion, so that the generated cross-scale feature vector can objectively reflect the actual operating state of the frequency converter, providing an accurate data basis for subsequent operating condition classification and control parameter setting.

[0029] 2. This invention uses a multi-task feedforward neural network containing classification and regression branches to synchronously output the operating condition classification results and the initial set values ​​of control parameters. Combined with the power supply mapping table, it formulates limiting voltage increase and power supply mode switching logic for spark precursor and back corona conditions. When the classification confidence verification is qualified and the equipment safety conditions are met, the target control parameters are updated, thereby realizing the adaptive adjustment of the output parameters of the frequency converter under different power supply modes.

[0030] 3. This invention constructs a deterministic boundary constraint mechanism that includes state triggering conditions and an independently operating asynchronous hardware protection logic. When the algorithm reasoning is abnormal or the classification confidence is insufficient, the power increase is limited by the software system. When an overcurrent occurs, the drive signal is directly cut off by a fast comparator and the hardware latch is maintained. This makes up for the potential risks brought about by the uninterpretability of deep learning algorithms and improves the overall safety of the frequency converter power supply control system. Attached Figure Description

[0031] Figure 1 This is a system architecture diagram of one embodiment of the present invention; Figure 2 This is a flowchart of a method according to an embodiment of the present invention; Figure 3 This is a flowchart of macroscopic working condition adaptive mapping according to an embodiment of the present invention; Figure 4 This is a timing diagram of high-frequency signal synchronous acquisition and ripple removal processing according to an embodiment of the present invention; Figure 5 This is a flowchart of gated modulation and feature fusion according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the constrained control execution principle according to an embodiment of the present invention; Figure 7 This is a comparison diagram of the evolution trajectory of the secondary voltage of the overcurrent protection system as described in this invention and that of the traditional overcurrent protection system. Detailed Implementation

[0032] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] See attached document Figure 1 This invention provides a method for regulating the operating conditions of a variable frequency power supply for an electrostatic precipitator (ESP) based on deep learning. The system for implementing this method includes a power supply circuit, a data acquisition module, and a control platform. The power supply circuit is located at the front end of the ESP in a thermal power plant and includes a rectifier unit, an inverter bridge, a step-up transformer, and a rectifier silicon stack connected in sequence. The rectifier unit converts the plant's AC power into a DC bus voltage. The inverter bridge outputs high-frequency AC power according to a pulse-width modulation (PWM) drive signal, which, after step-up and rectification, outputs high-voltage DC power or high-voltage pulsed power to the ESP.

[0034] The acquisition module is used to acquire real-time secondary electrical signals within the electric field, including a secondary voltage detection branch and a secondary current detection branch. The secondary voltage detection branch includes a high-voltage divider, an isolation amplifier circuit, a secondary voltage anti-aliasing filter circuit, and a secondary voltage analog-to-digital conversion channel. The high-voltage divider is connected between the rectifier silicon stack output terminal and the ground electrode, and its low-voltage detection terminal is connected to the isolation amplifier circuit. The analog secondary voltage signal output by the isolation amplifier circuit enters the secondary voltage analog-to-digital conversion channel after passing through the secondary voltage anti-aliasing filter circuit, generating the original high-frequency secondary voltage sampling signal.

[0035] The secondary current detection branch includes a current sensor, a signal conditioning circuit, a secondary current anti-aliasing filter circuit, a secondary current analog-to-digital converter (ADC) channel, and a fast comparator. The current sensor is placed in the secondary current path of the power supply loop. The first output of the signal conditioning circuit is connected to the secondary current ADC channel via the secondary current anti-aliasing filter circuit to generate a secondary current sampling signal. The second output of the signal conditioning circuit is connected to the fast comparator. The output of the fast comparator is connected to the hardware blocking terminal of the inverter bridge gate drive circuit and the hardware latch circuit. When the secondary current exceeds the hardware overcurrent protection threshold, the fast comparator directly outputs a hardware blocking signal, asynchronously cutting off the gate drive of the inverter bridge switching devices. The hardware latch circuit maintains the hardware latch blocking state after the hardware blocking signal is triggered.

[0036] In the following text, the instantaneous protection signal output by the fast comparator is called the hardware lockout signal, the lockout state maintained by the hardware latch circuit is called the hardware latch lockout state, and the software trigger state recorded by the control platform based on the hardware latch lockout state is called the hardware lockout trigger state. The reset enable output of the signal processor is connected to the reset control terminal of the hardware latch circuit, and the status output of the hardware latch circuit is connected to the protection status input of the signal processor. The hardware latch lockout state is then transmitted to the industrial computer via the signal processor or gate array chip. The status output of the fast comparator is also connected to the comparator status input of the signal processor, and the signal processor transmits the current output state of the fast comparator to the industrial computer.

[0037] The control platform includes a gate array chip, an industrial computer, and a signal processor. The gate array chip connects two analog-to-digital conversion channels and is used to buffer sampled data, append timestamps, and perform synchronization preprocessing. The signal processor outputs a carrier synchronization trigger signal to the gate array chip, which generates a logic-locked sampling phase through a phase-locked loop and a sampling clock, and provides the same sampling clock to both analog-to-digital conversion channels.

[0038] Both secondary voltage and secondary current sampling data are appended with sampling timestamps. The control platform prioritizes time alignment based on the sampling sequence number; if the sampling sequence numbers cannot be matched one-to-one, the nearest sampling timestamp is used for matching, and sampling points whose timestamp differences exceed the preset maximum allowable time deviation are marked as invalid.

[0039] The industrial control computer internally deploys an evaluation network and a control network, and communicates with the gate array chip via a high-speed serial bus. The signal processor receives control parameter commands from the industrial control computer and generates corresponding pulse width modulation drive signals based on the power supply mode, fundamental voltage amplitude, pulse power supply frequency, and pulse duty cycle. The fundamental voltage amplitude is used to determine the fundamental amplitude of the inverter bridge's high-frequency AC output; the pulse power supply frequency is used to determine the repetition frequency of the high-voltage pulse envelope; and the pulse duty cycle is used to determine the proportion of effective power supply time within one pulse power supply cycle. The intermittent power supply mode also uses preset intermittent on-time and preset intermittent off-time determined according to equipment safety tests; both are stored as fixed mode parameters and are not output by the control network's return branch.

[0040] The signal processor feeds back the three continuous control parameters and power supply mode actually used in the previous control update cycle to the industrial control computer. The three continuous control parameters serve as feedback control state values. In DC power supply mode, only the fundamental voltage amplitude is the current effective control component; in pulse power supply mode and intermittent power supply mode, all three feedback control state values ​​are effective control components. The industrial control computer inputs the cross-scale feature vector, feedback control state values, and current power supply mode into the control network to generate initial setpoints for the control parameters.

[0041] The system updates the macroscopic operating status and triggers the evaluation network at a second-level cycle, collects secondary voltage and secondary current at a microsecond-level cycle, and performs feature extraction, phase space reconstruction, network inference, boundary constraints, and command issuance at a millisecond-level time sliding window. Each preset sliding step forward of the time sliding window forms a control update cycle. The macroscopic operating status update cycle is longer than the control update cycle, and all types of operating data and control data are timestamped.

[0042] Upon initial system startup, the signal processor operates using preset safety initial control parameters. The control network is not activated until the first set of currently valid dynamic parameters, the valid phase ripple template, the required historical data, and the first valid finite-time maximum Lyapunov exponent are formed. The control platform stores the preset safety initial control parameters, the preset safety control parameters corresponding to each power supply mode, and the preset low-power pulse control parameter set. When switching to pulse power supply mode due to spark precursors, the preset low-power pulse control parameter set is designated as the mode-switching safety control parameters; when switching to other power supply modes, the preset safety control parameters corresponding to the candidate power supply mode are designated as the mode-switching safety control parameters.

[0043] After the power supply mode switch is completed, the mode switching safety control parameters are used for operation until the phase ripple template is reconstructed, the actual historical data length meets the requirements, and a valid cross-scale feature vector is re-formed. The completion of the new phase ripple template construction, the fulfillment of the actual historical data length requirements, and the re-formation of a valid cross-scale feature vector are collectively referred to as the template reconstruction and recovery condition being met. The preset low-power pulse control parameter group includes the preset low-power fundamental voltage amplitude, the preset low-power pulse power supply frequency, and the preset low-power pulse duty cycle. The low-power pulse operating state formed by this parameter belongs to the pulse power supply mode.

[0044] See attached document Figure 2 The main implementation steps of this method are as follows: S10: Obtain boiler load, dust collector inlet flue gas temperature, and flue gas flow rate, and input them into the evaluation network to form candidate delayed sampling point number, candidate embedding dimension, and candidate dynamic evolution threshold. These three constitute an inseparable candidate dynamic parameter set, which is then switched to the current effective dynamic parameter set after satisfying the boundary conditions of stability, historical data length, and control update cycle.

[0045] S20 synchronously acquires the original high-frequency secondary voltage sampling signal and secondary current sampling signal. Based on the phase residual of continuous effective historical switching cycles, the secondary voltage signal is template-cancelled to form a one-dimensional net voltage timing signal. The secondary current signal is filtered, downsampled, and time-aligned to form the secondary current mean, secondary current change rate, and voltage-current correlation characteristics.

[0046] S30: Based on the current number of effective delayed sampling points and the current effective embedding dimension, extract the net voltage sampling points within the same continuous effective data segment from the historical voltage data buffer queue, form a multi-dimensional phase space state vector, and assemble a multi-dimensional phase space trajectory matrix in chronological order.

[0047] S40: Based on the multidimensional phase space trajectory matrix, a finite-time maximum Lyapunov exponent input feature is formed, and the micro-feature gating coefficient is determined in combination with the current effective dynamic evolution threshold. The gating coefficient is used to fuse micro-electrical features and macro-operational state data to generate a cross-scale feature vector.

[0048] S50 inputs the cross-scale feature vector, the feedback control state value from the previous control update cycle, and the current power supply mode into the control network, and outputs the classification probabilities corresponding to three micro-conditions: normal corona, spark precursor, and anti-corona, as well as the initial setpoints for the fundamental voltage amplitude, pulse power supply frequency, and pulse duty cycle. The control platform determines the permissible action direction based on the micro-condition classification results obtained after confidence verification, and forms candidate power supply modes according to deterministic mode generation rules. The control network or deterministic mode generation rules generate initial setpoints for control parameters. After verification by single maximum change limits, absolute value limits, and mode switching conditions, the final control parameter setpoints are generated, then encapsulated and encoded using a communication protocol to form control parameter commands, which are then sent to the signal processor. The physical quantities actually used by the signal processor to generate the drive signals are the actual control parameters used.

[0049] See attached document Figure 3 In a specific embodiment, step S10 may include the following sub-steps: S110, the control platform reads boiler load, dust collector inlet flue gas temperature, and flue gas flow rate from the distributed control system of the thermal power plant at a set second-level cycle. These three data points together characterize the macroscopic operating conditions that affect flue gas volume, dust conductivity, and partial discharge evolution.

[0050] S120, the control platform standardizes the macroscopic input vector using the mean and standard deviation determined by the training set of the evaluation network. The evaluation network adopts a multilayer perceptron structure, including three input nodes, two hidden layers, and three output nodes, which output the predicted values ​​of delay time, embedding dimension, and dynamic evolution threshold, respectively.

[0051] Before being put into online operation, the control platform uses the historical operating database of the electrostatic precipitator to train and evaluate the network. The historical operating database includes at least boiler load, flue gas temperature at the inlet of the precipitator, flue gas flow rate, high-frequency secondary voltage signal, high-frequency secondary current signal, power supply mode, feedback control status value, outlet dust concentration, power metering data, hardware shutdown signal trigger records, and expert diagnosis records. The control platform divides the historical data into training set, validation set, and test set according to the operating date. The same hardware shutdown signal trigger event confirmed by expert diagnosis as being caused by spark discharge and its precursor time intervals do not span different datasets.

[0052] The control platform uses the same synchronous partitioning, template cancellation, filtering, downsampling, time alignment, window truncation, phase space reconstruction, and finite-time maximum Lyapunov exponent calculation rules as in the online operation phase to process historical signals. For the historical time windows in the training set, the average mutual information method is used to determine the number of delayed sampling points, and candidate labels for delayed time are formed accordingly; the pseudo-nearest neighbor method is used to determine candidate labels for embedding dimension. Spark precursor samples are selected from a preset precursor time interval before a hardware blocking signal trigger event confirmed by expert diagnosis as being caused by spark discharge. Normal corona samples are selected from the operating range where the interval between any hardware blocking signal trigger event and the event is not less than a preset safety time, and the secondary voltage and secondary current are within a preset stable range.

[0053] The control platform divides the training samples into multiple macroscopic operating condition sample groups according to the standardized macroscopic input vector. Sample groups with insufficient sample size are merged with the sample group whose center vector is closest. The median of the candidate delay time labels within the merged sample group is taken as the delay time label, and the mode of the candidate embedding dimension labels is taken as the embedding dimension label; when multiple modes exist, the smaller mode is used. The control platform uses the delay time label and the embedding dimension label to perform phase space reconstruction, calculates the maximum Lyapunov exponent for the effective finite time of normal corona samples and spark precursor samples, and uses the midpoint of adjacent different values ​​as a candidate threshold.

[0054] For any candidate threshold, windows with an exponent greater than or equal to the threshold are designated as candidate windows for spark precursors, while windows with an exponent less than the threshold are designated as candidate windows for normal corona discharges. The control platform selects the candidate threshold that satisfies the preset upper limit for the false alarm rate of normal corona discharges and has the lowest false alarm rate for spark precursors as the dynamic evolution threshold label; when the false alarm rates are the same, the threshold with the lower false alarm rate is selected; when both are the same, the threshold with the lower value is selected. If the number of valid exponent samples for any category is insufficient, if there are fewer than two different exponent values, or if there is no candidate threshold that satisfies the upper limit for the false alarm rate, the dynamic evolution threshold label is marked as invalid, and the training loss for the corresponding output is not calculated. After the three labels are generated, they serve as the supervision labels for the corresponding historical time windows.

[0055] The evaluation network's delay time output, embedding dimension output, and dynamic evolution threshold output are each calculated using mean squared error loss normalized to the output error scale. The total loss of the evaluation network is formed by combining the delay time output loss, embedding dimension output loss, and dynamic evolution threshold output loss according to preset weights. These preset weights are determined using the validation set and remain fixed after the evaluation network training is completed. The control platform uses Euclidean distance to calculate the distance between each standardized macroscopic input vector in the training set and the center vector of its corresponding sample group, and the preset quantile value of the aforementioned distance is determined as the out-of-distribution distance threshold. The evaluation network updates network parameters using the training set, selects network parameters and executes a stop-training judgment using the validation set, and evaluates performance using the test set. After training, the evaluation network structure, weight parameters, standardized parameters, predicted value clipping boundaries, sample group center vectors, and out-of-distribution distance thresholds are fixed.

[0056] During online operation, the control platform uses the same Euclidean distance to calculate the distance between the current standardized macroscopic input vector and the center vectors of each sample group. When the distance between the current macroscopic input vector and the center vector of the nearest sample group exceeds the out-of-distribution distance threshold, the out-of-distribution trigger state of the macroscopic input is recorded, the previous set of currently valid dynamic parameter groups is maintained, and new candidate dynamic parameter groups are not enabled; when the system starts for the first time and there is no previous set of currently valid dynamic parameter groups, the preset safe initial control parameters are maintained. After the current macroscopic input's macroscopic operating state update cycle returns to the training data distribution range for a consecutive preset number of times, the out-of-distribution trigger state is deactivated.

[0057] S130: The standardized macroscopic input vector is forward-propagated through the evaluation network to generate three predicted values. The number of candidate delayed sampling points is determined by rounding the ratio of candidate delay time to effective sampling period; the candidate embedding dimension is determined by integerization and range constraints; the candidate dynamic evolution threshold is determined by upper and lower limit pruning. Each pruning boundary is determined based on the training label range, the capacity of the historical voltage data buffer queue, and the allowable computational cost of the control update cycle, and remains fixed during online operation.

[0058] S140. The three candidate parameters formed in the same macroscopic operating state update are managed as an indivisible parameter group. If the number of candidate delayed sampling points and the candidate embedding dimension are the same in two consecutive macroscopic operating state update cycles, and the difference in the candidate dynamic evolution threshold does not exceed the preset threshold variation tolerance, the candidate dynamic parameter group is determined to meet the stability condition. If the candidate parameter group simultaneously meets the historical data length requirement and the control update cycle boundary condition, it is activated as a whole in the next control update cycle. The first group of candidate parameters is not required to meet the stability condition for two consecutive macroscopic operating state update cycles, but it should meet the phase ripple template construction requirement, the historical data length requirement, and the control update cycle boundary condition.

[0059] See attached document Figure 4In one specific embodiment, step S20 may include the following sub-steps: The S210 secondary voltage analog-to-digital conversion channel and the secondary current analog-to-digital conversion channel use the same sampling clock to generate the original high-frequency secondary voltage sampling signal and the secondary current sampling signal, respectively. The gate array chip synchronously receives the two types of sampling signals, adds a sampling timestamp, and buffers them.

[0060] S220: The signal processor outputs a carrier synchronization trigger signal to the gate array chip. The gate array chip uses a phase-locked loop to ensure that the sampling frequency of the analog-to-digital converter circuit is a fixed integer multiple of the pulse-width modulation carrier frequency, and marks the start position and phase sampling point number of each switching cycle. Under the same effective carrier frequency setting, the control network does not change the pulse-width modulation carrier frequency. When the carrier frequency changes due to equipment settings adjustments, the control platform marks the existing data in the current phase ripple template and historical voltage data buffer queue as invalid, resets the sampling frequency and the effective sampling period of the one-dimensional net voltage timing signal, and loads the evaluation network, control network, and their parameters associated with the new carrier frequency. If there is no corresponding effective carrier frequency setting, network inference is not performed, and the preset safety control parameters corresponding to the current power supply mode are used for operation.

[0061] S230, the control platform arranges the original high-frequency secondary voltage sampling signals according to the switching cycle and phase sequence, calculates the periodic average voltage for each historical switching cycle, and subtracts the periodic average voltage from each phase sampling value to obtain the phase residual. When the system first constructs the phase ripple template, it uses the current pulse width modulation ratio or equivalent driving parameters as the temporary template reference driving parameters; after the template is constructed, it saves the power supply mode, carrier frequency, and template reference driving parameters at the time of template construction.

[0062] A historical switching cycle is considered valid when the number of sampling points reaches a preset number, the sampled values ​​are within the effective range, there is no data loss, communication anomaly, hardware blocking signal, or effective hardware latch blocking state, and its power supply mode and carrier frequency are the same as the current switching cycle. In pulse power supply mode and intermittent power supply mode, both the historical switching cycle and the current switching cycle must be within the effective power supply range of the pulse envelope. When constructing the template for the first time, the difference between the driving parameters affecting the carrier ripple shape and the temporary template reference driving parameters must not exceed the template's applicable tolerance. If the template already exists, the difference between the driving parameters and the saved template reference driving parameters must not exceed the template's applicable tolerance. Ineffective power supply ranges of the pulse envelope and intermittent disconnection times are not included in template construction and create data continuity interruptions in the historical voltage data buffer queue.

[0063] The control platform determines the minimum continuous effective data duration based on the time sliding window, the maximum historical backtracking range required by the phase space state vector, and the maximum subsequent evolution range required for calculation using the maximum Lyapunov exponent over a finite time. The maximum historical backtracking range is determined by the number of delayed sampling points, the embedding dimension, and the effective sampling period, while the maximum subsequent evolution range is determined by the maximum number of evolution steps and the effective sampling period. When the time sliding window, the maximum historical backtracking range, and the maximum subsequent evolution range overlap, the overlapping portion is calculated only once. The minimum continuous effective data duration should ensure that time sliding window truncation, phase space state vector generation, and subsequent evolution distance tracking can be completed within the same continuous effective data segment. Before switching power supply modes, the theoretical continuous effective power supply duration is verified based on the candidate power supply mode and its corresponding target control parameters; after switching power supply modes, it is verified whether the same continuous effective data segment in the historical voltage data buffer queue has reached the required duration. The continuous effective power supply interval in the pulse power supply mode is determined based on the pulse power supply frequency and pulse duty cycle, while the intermittent power supply mode is also subject to a preset intermittent on-time limitation. When the minimum continuous effective data duration requirement cannot be met, a new finite-time maximum Lyapunov exponent is not calculated, and a new cross-scale feature vector is not formed. When in the template reconstruction stage after power supply mode switching, the mode switching safety control parameters are used for operation. In other cases, the preset safety control parameters corresponding to the current power supply mode are used for operation.

[0064] The template tolerance is determined through equipment testing. Under the condition that the power supply mode, carrier frequency and fixed drive settings remain unchanged, the pulse width modulation ratio or equivalent drive parameters are changed and the deviations between the corresponding phase ripple templates are compared. The maximum allowable difference in drive parameters when the template deviation does not exceed the preset ripple cancellation error upper limit is determined as the template tolerance.

[0065] If the number of sampling points in the current switching cycle is insufficient, the sampled value exceeds the effective range, the sampled data is lost, the analog-to-digital conversion channel is abnormal, a hardware blocking signal is triggered, or the hardware latch blocking state is valid, the secondary voltage data for that cycle will be marked as invalid. This cycle's data will not be used to generate a net voltage sampled value, nor will it be written to the historical voltage data buffer queue. If the common sampling clock, synchronous trigger, or shared communication link is abnormal, the synchronous secondary current sampling points will be marked as invalid. If only the secondary voltage branch is abnormal and the secondary current data is valid, the synchronous secondary current data will be retained for calculating the average secondary current and the rate of change of secondary current, but the voltage-current correlation characteristics will be marked as invalid.

[0066] ; When the number of consecutive effective historical switching cycles reaches the template construction requirement, the control platform calculates the arithmetic mean of the phase residuals with the same phase number in each historical switching cycle to form a phase ripple template: The control platform subtracts the phase ripple template value of the corresponding phase from the current original high-frequency secondary voltage sample value to form the net voltage sample value: ; In the formula, Indicates the current pulse width modulation switching cycle number. This indicates the sequence number of the phase sampling point within one pulse width modulation switching cycle. This indicates the number of consecutive valid historical switching cycles used to construct the phase ripple template. An index variable representing a continuous valid historical switching cycle. Indicates the first Before the pulse width modulation switching cycle, the first Within the first consecutive effective historical switching cycle, the first The original high-frequency secondary voltage sampled values ​​corresponding to each phase sampling point Indicates the first Before the pulse width modulation switching cycle, the first The periodic average voltage corresponding to a consecutive effective historical switching cycle Indicates the first The phase ripple template value corresponding to each phase sampling point Indicates the first Within the pulse width modulation switching cycle, the first The original high-frequency secondary voltage sampled values ​​corresponding to each phase sampling point Indicates the first Within the pulse width modulation switching cycle, the first The net voltage sample value corresponding to each phase sampling point.

[0067] The phase ripple template is not updated when the number of consecutive valid historical switching cycles is insufficient. If a previous valid phase ripple template exists, it continues to be used; otherwise, the safety control parameters determined for the current stage are used. If the cycle average voltage changes abruptly but the template's applicable conditions remain unchanged, template updates are paused and the previous valid template continues to be used. If the power supply mode, carrier frequency, or fixed drive settings change, or the drive parameters exceed the template's applicable range, the current template and existing historical data are marked as invalid. During the template reconstruction phase after a power supply mode switch, the drive parameters corresponding to the mode switching safety control parameters are used as the temporary template reference; otherwise, the drive parameters corresponding to the preset safety control parameters of the current power supply mode are used as the temporary template reference to reconstruct the phase ripple template.

[0068] S240: The net voltage sample value is expanded according to the sampling time, and after digital low-pass filtering and downsampling by a preset integer multiple, a one-dimensional net voltage timing signal is formed. The cutoff frequency of the digital low-pass filter is lower than half of the sampling frequency after downsampling, and it covers the frequency band of corona discharge variation characteristics that need to be retained.

[0069] S250: The secondary current sampling signal, after digital low-pass filtering and downsampling by the same factor, is aligned with the one-dimensional net voltage timing signal according to the sampling sequence number or timestamp. The control platform calculates the mean value of the secondary current within the same time sliding window as the phase space reconstruction, using the slope of the first-order linear fit between the secondary current sampling value and the sampling time as the rate of change of the secondary current, and the zero-delay Pearson linear correlation coefficient as the voltage-current correlation feature.

[0070] When the variance of the net voltage or secondary current within the current window is lower than the preset minimum variance, the voltage-current related features are marked as invalid. When there is a communication error in the secondary current channel, insufficient effective sampling points, or the range is exceeded, the three secondary current related features are marked as invalid. When the fitting or correlation coefficient calculation fails, the corresponding feature is marked as invalid. The control platform sets invalid trigger sub-states for the three features respectively. If the number of consecutive invalid cycles of a feature does not exceed the preset number of cycles to be used and there is a most recent valid value, the valid value is used; otherwise, the corresponding safe replacement value is used, where the safe replacement values ​​for the voltage-current related features and the rate of change of secondary current are both 0, and the mean value of the secondary current is the median of the normal corona safe operation samples in the training set. When any sub-state is valid, an invalid trigger state for the secondary current related features is established and the increase of the fundamental voltage amplitude is prohibited; after the corresponding feature is valid for a consecutive preset number of recovery cycles, its sub-state is released, and after all sub-states are released, the overall trigger state is released.

[0071] In a specific embodiment, step S30 above may specifically include the following sub-steps: S310: The control platform constructs a historical voltage data buffer queue according to the first-in, first-out (FIFO) principle, storing the net voltage sample value, sampling timestamp, and data continuity flag. Invalid sampling periods, ineffective power supply intervals in the pulse envelope, intermittent disconnection times, and abnormal sampling time intervals all cause data continuity interruptions. Valid data length is only calculated within the same continuous valid data segment. If the required historical data length for a candidate dynamic parameter group exceeds the current valid data length, the previous currently valid dynamic parameter group is used until the data length meets the requirement.

[0072] S320: The control platform sets a time sliding window on the historical voltage data buffer queue and extracts the valid data segment of the current control update cycle according to the preset sliding step size.

[0073] S330, for the target net voltage sampling point within the time sliding window, takes this sampling point as the end point and selects the current effective embedding dimension minus one historical sampling point in the historical direction according to the current effective delayed sampling point number, forming a multi-dimensional phase space state vector. State vectors that need to cross data continuity interruptions are marked as invalid. The control platform arranges the effective state vectors in chronological order to form a multi-dimensional phase space trajectory matrix.

[0074] See attached document Figure 5 In a specific embodiment, step S40 may include the following sub-steps: S410, the finite-time maximum Lyapunov exponent calculation result is the original result obtained from the phase space trajectory in the current control update cycle; the result verified by the number of nearest neighbors, fitting interval, goodness of fit and calculation time limit is the effective finite-time maximum Lyapunov exponent; after the effective value is adopted, the most recent effective value is carried over or the risk substitution value is processed, the value actually used for gating calculation and cross-scale feature fusion is the finite-time maximum Lyapunov exponent input feature.

[0075] The control platform selects a finite number of reference state vectors from the multidimensional phase space trajectory matrix. After excluding candidate state vectors that are too close in time, it selects the nearest neighbor. It tracks the distance change between the reference state vector and the nearest neighbor over multiple evolutionary steps, takes the natural logarithm of the distance and averages it. The slope is calculated from the continuously increasing interval with the required goodness of fit, and the result of the finite-time maximum Lyapunov exponent is obtained based on the actual time interval of the evolutionary steps. If there are insufficient effective nearest neighbors, no effective fitting interval, or the calculation exceeds a preset time limit, the result is marked as invalid. When the finite-time maximum Lyapunov exponent calculation result is marked as invalid, the control platform records the invalid finite-time maximum Lyapunov exponent trigger state; after obtaining a valid finite-time maximum Lyapunov exponent and meeting the recovery conditions specified in S530, the trigger state is released.

[0076] The control platform statistically analyzes the median of the maximum effective finite-time Lyapunov exponents corresponding to spark precursor samples in the control network training set, and sets this median as the preset risk benchmark. When the number of effective spark precursor samples is lower than the preset minimum risk sample number, the maximum value of the effective dynamic evolution threshold label in the evaluation network training set is used as the preset risk benchmark. When there is no effective dynamic evolution threshold label in the evaluation network training set, the risk substitution value processing path is not enabled. When risk substitution values ​​need to be generated, the control platform compares the preset risk benchmark value with the current effective dynamic evolution threshold, and determines the larger value as the risk substitution value. When the current result is valid, it is used as the input feature of the maximum Lyapunov exponent over a finite time period and the number of consecutive invalid periods is cleared to zero. When the result is invalid but the number of consecutive invalid periods has not reached the preset risk period number, the most recent valid value is used. If there is no most recent valid value, the corresponding safety control parameter is adopted. When the number of consecutive invalid periods reaches the preset risk period number and the risk substitution path is available, the risk substitution value is adopted, the micro-feature gating coefficient is directly set to the preset maximum gating value, and the increase of the fundamental voltage amplitude is prohibited. When the risk substitution path is unavailable, the control network output is not adopted, and it is gradually adjusted to the preset safety control parameter corresponding to the current power supply mode.

[0077] In S420, except for the case in S410 where the micro-feature gating coefficients are directly set to the preset maximum gating value, the control platform uses fixed gating standardization parameters from the training phase. It performs the same standardization process on the input feature with the maximum Lyapunov exponent over a finite time and the current effective dynamic evolution threshold. The micro-feature gating coefficients are determined based on the difference between the two using a monotonically increasing Sigmoid function with values ​​ranging from zero to one. The micro-feature gating coefficients are limited to a preset minimum and a preset maximum value, and under normal conditions, the maximum change between two adjacent control update cycles is also limited.

[0078] The S430 control platform uses standardized feature parameters determined by the control network training set and fixed after training to standardize the finite-time maximum Lyapunov exponent input feature, the mean secondary current, the rate of change of the secondary current, the voltage-current correlation feature, the boiler load, the dust collector inlet flue gas temperature, and the flue gas flow rate. The control platform multiplies the standardized finite-time maximum Lyapunov exponent input feature, the standardized mean secondary current, the standardized rate of change of the secondary current, and the standardized voltage-current correlation feature by a micro-feature gating coefficient; it multiplies the standardized boiler load, the standardized dust collector inlet flue gas temperature, and the standardized flue gas flow rate by one minus the micro-feature gating coefficient. The control platform then concatenates these features in the following order: standardized finite-time maximum Lyapunov exponent input feature, standardized mean secondary current, standardized rate of change of the secondary current, standardized voltage-current correlation feature, standardized boiler load, standardized dust collector inlet flue gas temperature, and standardized flue gas flow rate, forming a seven-dimensional cross-scale feature vector. The same feature arrangement order is used in both the offline training and online operation phases.

[0079] See attached document Figure 6 In a specific embodiment, step S50 may specifically include the following sub-steps: The S510 control network employs a multi-task feedforward neural network, including a shared input layer, a shared hidden layer, a classification branch, and a regression branch. The shared hidden layer comprises a first and a second shared hidden layer, both fully connected sequentially, and uses a linear rectified activation function. The second shared hidden layer connects the classification and regression branches. The classification branch uses a normalized exponential function to output the classification probabilities corresponding to three microscopic operating conditions: normal corona, spark precursor, and anti-corona. The regression branch uses a linear output layer, outputting the fundamental voltage amplitude, pulse power supply frequency, and initial pulse duty cycle settings in physical quantity units. The current power supply mode uses three-dimensional one-hot encoding. When the cross-scale feature vector is seven-dimensional, the control network input includes seven-dimensional cross-scale features, three feedback control state values, and the three-dimensional power supply mode encoding.

[0080] The control network training employs the same data processing and feature generation rules as the online operation. Historical time windows located within a preset precursor time interval before the occurrence of a hardware blocking signal trigger event confirmed by expert diagnosis as a spark discharge are marked as spark precursors. The control platform uses historical time windows confirmed by expert diagnosis as reverse corona that do not fall within the spark precursor time interval as reverse corona reference samples, and historical time windows where the secondary voltage and secondary current are within a preset stable range, the interval between any hardware blocking signal trigger event and the event is not less than a preset safe time, and which have not been diagnosed by experts as reverse corona as normal corona reference samples. Based on the reverse corona reference samples and the normal corona reference samples, the control platform evaluates different judgment directions and candidate thresholds for the average secondary current, the rate of change of secondary current, and voltage-current correlation characteristics. Within the preset upper limit of the normal corona false alarm rate, the platform selects the judgment direction and threshold with the lowest reverse corona false alarm rate as the reverse corona judgment condition for the corresponding feature. For historical time windows that do not fall within the spark precursor time interval, if they are confirmed by experts to be inverse corona discharge and at least two of the three characteristics meet the corresponding inverse corona discharge criteria within a consecutive preset number of historical time windows, they are marked as inverse corona discharge. Other historical time windows that meet the criteria for normal corona discharge baseline samples are marked as normal corona discharge. When the same historical time window simultaneously meets multiple classification criteria, a unique classification label is determined according to the order of spark precursor, inverse corona discharge, and normal corona discharge.

[0081] The control platform standardizes the cross-scale feature vectors and feedback control state values ​​using standardized parameters determined by the control network training set, and concatenates the standardized cross-scale feature vectors and standardized feedback control state values ​​into a historical input state vector. The control platform uses Euclidean distance to calculate the distance between historical input state vectors, identifying historical samples with the same power supply mode and a distance not exceeding a preset similarity threshold as similar historical samples. The preset similarity threshold is determined based on the preset quantile of the distance between each historical sample in the training set and its nearest historical sample with the same power supply mode. The regression labels for the control network are determined from the aforementioned similar historical samples. After the actual control parameters take effect within the historical time window and after a flue gas transport delay determined based on the effective flue volume and flue gas flow rate, the outlet dust concentration, the number of hardware blockade signal triggers, the energy consumption per unit flue gas volume, and the secondary voltage stability are evaluated within a preset effect evaluation time. Samples whose control parameters, macroscopic operating states, or other electric field states change beyond a preset tolerance during the evaluation period do not generate regression labels. Candidate control parameters with lower outlet dust concentrations are prioritized. When the outlet dust concentration difference does not exceed a preset tolerance, candidate control parameters with fewer hardware blocking signal triggers, lower power consumption, smaller secondary voltage fluctuations, and smaller changes relative to the feedback control state value are selected sequentially. Cross-entropy loss is used for the classification branch, and mean squared error loss normalized to the error scale is used for the regression branch. The control network updates its parameters using the training set, selects network parameters using the validation set, and executes a stop-training decision. The test set is used only for performance evaluation. After training, the control network structure, weight parameters, and standardized parameters are fixed, and the control network is not retrained during online operation.

[0082] The control platform establishes a power supply mapping table based on the frequency converter power supply equipment tests, recording the correspondence between the effective control components and the actual average power supply under each power supply mode. Within the allowable absolute value range of the equipment, the control platform sets multiple test nodes and determines the actual average power supply based on the energy metering data during the stable operation of each test node. In DC power supply mode, the control platform uses one-dimensional linear interpolation to determine the average power supply based on the adjacent fundamental voltage amplitude test nodes. In pulse power supply mode and intermittent power supply mode, the control platform uses multilinear interpolation to determine the average power supply based on the vertex data of each vertex of the test grid cell containing the control parameter to be queried. If the control parameter to be queried exceeds the coverage range of the test node, the test grid cell is incomplete, or the interpolation result is invalid, the power mapping result is marked as invalid. When the power mapping result is invalid, control actions or power supply mode switching based on the increase or decrease of the average power supply are not executed.

[0083] During online operation, the S520 classifies three probabilities and determines candidate micro-operating condition categories. When the maximum classification probability is lower than a preset confidence threshold, an insufficient classification confidence trigger state is recorded, no valid micro-operating condition classification result is generated, pattern generation and action direction judgment are not performed, no new instructions are generated using the current control network output, and the control parameter instructions from the previous control update cycle are retained. When the maximum classification probability reaches the preset confidence threshold, the candidate micro-operating condition category is determined as the micro-operating condition classification result.

[0084] In normal corona state, control parameters are allowed to perform steady-state optimization within deterministic boundary constraints. The spark precursor state records the spark precursor classification results and trigger states, prohibits increasing the fundamental voltage amplitude and average power supply, and allows decreasing the fundamental voltage amplitude, decreasing the pulse duty cycle, entering a low-power pulse operation state in pulse power supply mode, or switching from other power supply modes to pulse power supply mode and using a preset low-power pulse control parameter group. In reverse corona state, control actions such as decreasing the average power supply, decreasing the pulse duty cycle, or entering intermittent power supply mode are prioritized.

[0085] When the micro-condition classification result changes from normal corona to spark precursor or reverse corona for the first time, the control platform records the actual power supply mode and control parameters used before the change, forming the pre-abnormal power supply state. If the micro-condition classification result continuously recovers to normal corona, all safety trigger states are released, hardware latch blocking states are released, and the pre-abnormal power supply mode still falls within the equipment's allowed power supply mode, the pre-abnormal power supply mode and the actual control parameters used before the abnormality are used as the recovery target. If the pre-abnormal power supply mode differs from the current power supply mode, the pre-abnormal power supply mode is identified as a candidate power supply mode, and the corresponding mode switching safety control parameters are used to complete mode recovery. After the template reconstruction recovery conditions are met, parameter recovery begins. If the pre-abnormal power supply mode is the same as the current power supply mode, no power supply mode switching is performed; parameter recovery begins with the currently used control parameters. During parameter recovery, the system gradually recovers to the actual control parameters used before the abnormality according to the single maximum change limit, and clears the pre-abnormal power supply state after reaching the preset recovery tolerance.

[0086] When a power supply mode switch is triggered by a spark precursor or back corona state, the classification results should continuously reach the preset mode switching cycle number, the hardware latch blocking state should be released, and the candidate power supply mode and its target control parameters should meet the theoretical continuous effective power supply duration requirements. In the spark precursor state, if the current power supply mode is pulsed, the preset low-power pulse control parameter group is used as the target control parameter; if the current power supply mode is not pulsed, and the average power supply power of the preset low-power pulse control parameter group is not higher than the current average power supply power, the pulsed power supply mode is determined as the candidate power supply mode, the preset low-power pulse control parameter group is determined as the target control parameter, and a request to switch to the pulsed power supply mode is generated. In the back corona state, if the current power supply mode is not intermittent, and the average power supply power of the preset safety control parameters corresponding to the intermittent power supply mode is lower than the current average power supply power, a request to switch to the intermittent power supply mode is generated, and the preset safety control parameters corresponding to the candidate power supply mode are used as the target control parameter.

[0087] The control platform compares the initial setpoints of the control parameters with the feedback control state values ​​and determines the direction of the control action based on the power supply mapping table. When the control action direction is inconsistent with the classification result and the regression branch output, the inconsistent control action direction trigger state is recorded. The current control network output is not used, no mode switching request is generated, and the control parameter instructions from the previous control update cycle are retained. Similarly, when the current macroscopic input is outside the training data distribution or the system is in the phase ripple template reconstruction phase after a mode switch, the control network is not used to generate new control parameter instructions. After the power supply mode switch is completed, the control platform does not generate a new power supply mode switch request until the template reconstruction and recovery conditions are met.

[0088] S530, the control platform sets deterministic boundary constraints independent of the control network in the control parameter distribution link. Before entering the single maximum change limit, the control platform performs finite value judgment and physical quantity convertibility judgment on all target control parameters used in this control update cycle. Target control parameters include the control network regression branch output, the preset low-power pulse control parameter group, the preset safety control parameters corresponding to the candidate power supply mode, the recovery target control parameters in the power supply state before the anomaly, and the target control parameters corresponding to the external reset command. When any effective control component is non-numerical data, infinite data, or cannot be converted into a valid physical quantity, the control platform does not form a new control parameter setting value, records the abnormal trigger state of the control parameter, and retains the control parameter command of the previous control update cycle; target control parameters that can be converted into valid physical quantities but exceed the absolute value range allowed by the device are pruned by subsequent absolute value limits. When the control network inference exceeds the preset time limit, the control network output is marked as invalid, the control network inference timeout trigger state is recorded, the power supply mode switching request of this control update cycle is canceled, and the control parameter command of the previous control update cycle is retained. When no power supply mode switching request is generated, the control platform verifies all valid output components corresponding to the current power supply mode; when a power supply mode switching request is generated, it verifies all valid output components corresponding to the candidate power supply mode. If any verified output component contains non-numerical data, infinitely large data, or exceeds the output range allowed by the control network, the control network output is marked as invalid, the abnormal triggering state of the control parameters is recorded, the power supply mode switching request for this control update cycle is revoked, and the control parameter commands of the previous control update cycle are maintained.

[0089] Safety trigger states include: invalid secondary current-related characteristics trigger state, invalid maximum Lyapunov exponent trigger state, insufficient classification confidence trigger state, inconsistent control action direction trigger state, macroscopic input distribution trigger state, spark precursor classification result trigger state, control network inference timeout trigger state, control parameter anomaly trigger state, deterministic boundary constraint verification timeout trigger state, and hardware block trigger state. When any safety trigger state is active, increasing the fundamental voltage amplitude is prohibited. In pulse power supply mode and intermittent power supply mode, the average power supply power corresponding to the control parameter combination should not exceed the average power supply power corresponding to the actual control parameters used in the previous control update cycle. Each safety trigger state is released independently: when all secondary current-related feature invalid trigger sub-states are released, the corresponding overall trigger state is released; when the maximum Lyapunov exponent with a valid finite time is regained, the classification probability recovers to the confidence threshold, the classification result and control action direction are consistent, the macroscopic input is restored to the training distribution range, the classification result is no longer a spark precursor, the control network inference and boundary verification are restored to within the preset time limit, or the target control parameter is restored to validity, the corresponding trigger state is released after the corresponding conditions are continuously satisfied for their respective preset recovery cycle numbers. The release of the spark precursor classification result trigger state also requires the hardware latch blocking state to be invalid; when the hardware latch blocking state is released, the hardware blocking trigger state is released. The trigger state outside the macroscopic input distribution is counted according to the macroscopic operating state update cycle, and other states are counted according to the control update cycle. After all safety trigger states are released, the safety state flag prohibiting the increase of the fundamental voltage amplitude is released.

[0090] When there is no power supply mode switching request, the DC power supply mode only applies single maximum change limit and absolute value limit to the fundamental voltage amplitude; the pulse power supply mode and intermittent power supply mode apply the aforementioned limits to the fundamental voltage amplitude, pulse power supply frequency, and pulse duty cycle. When there is a power supply mode switching request, the target control parameter corresponding to the candidate power supply mode is used as the boundary constraint object; when there is no effective feedback control state value corresponding to the candidate power supply mode, the preset safety control parameter corresponding to the candidate power supply mode is used as the change limit benchmark.

[0091] The power supply mode switching request should also meet the following requirements: micro-condition classification results, classification confidence level, number of consecutive cycles of classification results being maintained, release of hardware latch blocking status, equipment safety conditions corresponding to the candidate power supply mode, and theoretical continuous effective power supply duration requirements. When all conditions are met, the candidate power supply mode and the target control parameters that have undergone variation and absolute value limitations form the control parameter setpoints; if any condition is not met, the power supply mode switching request is cancelled, the current power supply mode and the control parameter instructions of the previous control update cycle are maintained, and no new control parameter instructions are generated in this control update cycle.

[0092] The control parameter setpoints are encapsulated, converted to units, and encoded using communication protocols to form control parameter instructions. When the deterministic boundary constraint verification exceeds a preset time limit, the verification result of the current control update cycle is discarded, the control parameter instructions of the previous control update cycle are retained, and the deterministic boundary constraint verification timeout trigger status is recorded.

[0093] In S540, the signal processor generates a pulse-width modulation drive signal based on control parameter instructions that have passed deterministic boundary constraint verification. During the software control flow, the hardware asynchronous protection always runs independently and in parallel, with its protection priority higher than control network inference, deterministic boundary constraint verification, and software control parameter instructions.

[0094] The S550 signal processor feeds back the fundamental voltage amplitude, feedback pulse power supply frequency, feedback pulse duty cycle, current power supply mode, and hardware latch lockout status to the industrial control computer for use in the next control update cycle and subsequent offline training sample updates. Upon receiving valid feedback on the hardware latch lockout status, the control platform records the hardware lockout trigger state, cancels any pending power supply mode switching requests, and prohibits the generation of new control parameter commands via the control network. Simultaneously, it marks the current phase ripple template as invalid and writes a data continuity interruption flag to the historical voltage data buffer queue, ensuring that data before and after the hardware lockout signal trigger cannot form the same continuous valid data segment. After the hardware latch lockout status is released, the phase ripple template is reconstructed according to the preset safety control parameters or mode switching safety control parameters corresponding to the current power supply mode, and continuous valid historical data is accumulated. When the secondary current exceeds the hardware overcurrent protection threshold, the fast comparator directly outputs a hardware lockout signal, and the hardware latch circuit maintains the hardware latch lockout status. When the secondary current recovers to below the preset reset threshold, after a preset deionization waiting time, and the number of consecutive hardware lockout signal triggers within the preset time does not reach the fault count limit, the signal processor outputs a reset enable signal.

[0095] When the hardware latch lockout is released and there is no external mode change instruction, the signal processor maintains the power supply mode before the lockout and restores power according to the preset boost slope, starting from the preset safety control parameters corresponding to that mode. During the restoration period, each effective control component must not exceed the absolute value range allowed by the device, and the average power supply must not be higher than the average power supply before the hardware lockout signal was triggered; before the template reconstruction restoration conditions are met, no new control parameter instructions are generated using the control network.

[0096] When the manual reset instruction or the reset instruction from the superior control system requires changing the power supply mode, the mode switching is performed after the hardware latch blocking state is released, the fast comparator is in the unblocked state, the candidate power supply mode and its preset safety control parameters are valid and meet the requirements for theoretically continuous effective power supply duration; before the template reconstruction recovery condition is satisfied, the mode switching safety control parameters are used for operation. When the number of consecutive hardware blockage signal triggers within the preset time reaches the upper limit of fault times, the control platform maintains the inverter bridge blocking state and outputs a fault alarm. The deep learning algorithm cannot shield, delay or overwrite the hardware blockage signal.

[0097] Specific Application Embodiment: To further clarify the specific implementation process of the present invention, description is given by taking a variable-frequency power supply for an electric precipitator matched with a 600MW unit of a thermal power plant as an example. The following operating parameters and processing results are derived from the operation records of the variable-frequency power supply; the data used for effect comparison adopt statistical results under the same equipment conditions and similar macroscopic operating conditions.

[0098] It is set that the pulse width modulation carrier frequency of the variable-frequency power supply is 50kHz, and the corresponding switching period is 20μs; the sampling frequency of the analog-to-digital conversion channel is 5MHz, each switching period includes 100 phase sampling points, and the sequence numbers of the phase sampling points are integers between 1 and 100.

[0099] (1) Macro working condition adaptive mapping and parameter generation: The control platform reads the current macroscopic operating conditions from the distributed control system at a set second-level cycle: the boiler load is 450MW, the flue gas temperature at the precipitator inlet is 135°C, and the flue gas flow rate is 2.2×10 6 m 3 / h.

[0100] After standardizing the above data and inputting it into the evaluation network, the current candidate dynamic parameter group is formed through forward propagation output. After satisfying the stability condition of two consecutive macroscopic operating state update cycles, the overall switching is performed to the current effective dynamic parameter group: Current effective number of delayed sampling points: 3; Current effective embedding dimension: 4; Current effective dynamic evolution threshold: 0.45; Wherein, the current effective number of delayed sampling points acts on the one-dimensional net voltage timing signal formed after digital low-pass filtering and downsampling, and does not directly act on the original high-frequency secondary voltage sampling signal of 5MHz. The effective sampling period of the one-dimensional net voltage timing signal adopts the preset value corresponding to the current carrier frequency setting.

[0101] (2) Numerical calculation of signal processing and phase ripple template cancellation The system extracts the latest = 5 consecutive effective historical switching periods ( =1,2,3,4,5), used to construct the phase ripple template. The number of sampling points, sampling range, data integrity, communication status, hardware protection status, power supply mode, carrier frequency, and drive parameters of the five historical switching cycles all meet the valid historical switching cycle conditions specified in step S230. The five consecutive valid historical switching cycles are only used to illustrate the construction process of the phase ripple template; subsequent phase space reconstruction and finite-time maximum Lyapunov exponent calculation use the consecutive valid data segments in the historical voltage data buffer queue that meet the minimum consecutive valid data duration requirement.

[0102] The phase residuals at the 10th phase sampling point for the five historical switching cycles are 1.2kV, 1.3kV, 1.2kV, 1.3kV, and 1.0kV, respectively. The sum of the five phase residuals is 6.0kV. After calculating the arithmetic mean, the phase ripple template value corresponding to the 10th phase sampling point is 1.2kV.

[0103] Within the current pulse width modulation switching cycle, the original high-frequency secondary voltage sample value corresponding to the 10th phase sampling point is 61.6kV. After subtracting the 1.2kV phase ripple template value corresponding to this phase, the net voltage sample value is 60.4kV. The control platform processes the other phase sampling points within the current switching cycle in the same way, and forms a one-dimensional net voltage timing signal through digital low-pass filtering and downsampling.

[0104] Template cancellation reduces the periodic ripple components related to the carrier phase and forms a one-dimensional net voltage timing signal for subsequent phase space reconstruction.

[0105] (3) Explanation of finite-time maximum Lyapunov exponent processing and gating fusion: Using the current effective number of delayed sampling points (3) and the current effective embedding dimension (4), a time sliding window is extracted from the historical voltage data buffer queue to construct a multidimensional phase space trajectory matrix. The system tracks the logarithmic slope of the distance change between the reference state vector and the nearest neighbor within multiple evolution steps. The calculation result is valid, and the current finite-time maximum Lyapunov exponent calculation result is obtained as 0.65, which is used as the input feature for the finite-time maximum Lyapunov exponent.

[0106] The control platform uses fixed gating and standardization parameters from the training phase to perform the same standardization process on the input feature with the maximum Lyapunov exponent over a finite time and the current effective dynamic evolution threshold. The micro-feature gating coefficient is then calculated using a monotonically increasing Sigmoid function determined during the training phase, based on the difference between the two. In this embodiment, the input feature with the maximum Lyapunov exponent over a finite time is 0.65, the current effective dynamic evolution threshold is 0.45, and the calculated micro-feature gating coefficient is approximately 0.5498.

[0107] The control platform also obtains the average secondary current, rate of change of secondary current, and voltage-current correlation characteristics from a sliding window at the same time. Using the standardized parameters fixed during the control network training phase, it standardizes these three characteristics, as well as the boiler load, dust collector inlet flue gas temperature, and flue gas flow rate. The control platform processes the four micro-features and three macro-features according to the gating method specified in step S430, and concatenates them in a prescribed order to form a seven-dimensional cross-scale feature vector. Since the standardized parameters of each feature are fixed equipment settings after training, their specific values ​​are not repeated in this application embodiment.

[0108] (4) Control network reasoning and constraint enforcement: At the start of this control update cycle, the frequency converter is in DC power supply mode. The signal processor feeds back the fundamental voltage amplitude actually used in the previous control update cycle, as well as the saved candidate state values ​​of pulse power supply frequency and pulse duty cycle to the industrial control computer. The control platform processes the three feedback control state values ​​using standardized parameters fixed during the training phase, and inputs the seven-dimensional cross-scale feature vector, standardized feedback control state values, and current power supply mode encoding into the control network. The classification probabilities of normal corona, spark precursor, and anti-corona output by the classification branch are 0.12, 0.85, and 0.03, respectively. The classification probability corresponding to spark precursor is the highest and not lower than the preset confidence threshold; therefore, the micro-condition classification result for this control update cycle is determined to be spark precursor. After this classification result continuously reaches the preset number of mode switching cycles, the power supply mode switching condition verification begins.

[0109] The system records the trigger status of the spark precursor classification results and prohibits increasing the fundamental voltage amplitude and average power supply. After the spark precursor classification results continuously reach the preset mode switching cycle number, the control platform determines the pulse power supply mode as the candidate power supply mode and the preset low-power pulse control parameter group as the target control parameters. Verification shows that the current hardware latch blocking state is invalid, the power mapping result is valid, the average power supply corresponding to the preset low-power pulse control parameter group is not higher than the current average power supply, and the candidate power supply mode and its target control parameters meet the theoretical continuous effective power supply duration requirement. Therefore, a power supply mode switching request is generated. The preset low-power pulse control parameter group includes a fundamental voltage amplitude of 48kV, a pulse power supply frequency of 40Hz, and a preset low-power pulse duty cycle determined according to equipment safety tests.

[0110] The target control parameters are verified by limiting the maximum change in a single operation, limiting the absolute value, and checking the power supply mode switching conditions to form control parameter setpoints. These setpoints are then encapsulated, converted to units, and encoded using a communication protocol to form control parameter instructions, which are then sent to the signal processor. The signal processor generates pulse width modulation drive signals based on the control parameter instructions and feeds back the actual control parameters and the current power supply mode to the industrial control computer.

[0111] Comparison of running results: Both the control and experimental groups used the same electric field, the same frequency converter, the same sampling system, the same hardware overcurrent protection threshold, and the same dust concentration detection device. Statistical analysis was conducted during operating periods where the boiler load, dust collector inlet flue gas temperature, and flue gas flow rate were within similar ranges. Both groups used the same statistical period and data validity screening rules. The experiment was divided into two groups: Control group (traditional overcurrent protection method): passive power-off protection is achieved solely by relying on the hardware blocking signal output by the fast comparator after the secondary current exceeds the threshold.

[0112] Experimental group (method of this invention): Active regulation is implemented by deploying a control network.

[0113] See attached document Figure 7 ,like Figure 7 As shown by the solid black line, the traditional overcurrent protection method in the control group only triggers hardware latch blocking at the moment the channel discharge evolves to breakdown, causing a sharp drop in the one-dimensional net voltage timing signal. This not only generates extremely low residual voltage but also causes severe data continuity interruption; conversely, as Figure 7 As shown by the dark gray dashed line, the experimental group's method proactively identified the discharge risk at the pre-spark stage and immediately performed a mode switch, actively reducing the voltage and switching to a low-power pulse mode. This proactive regulation effectively avoided discharge breakdown, allowing the voltage waveform to smoothly transition to a safe pulse envelope state.

[0114] Experimental data statistics: Hardware blocking signal trigger frequency: Under the same statistical conditions, the control group experienced an average of 47 triggers per day, while the experimental group experienced an average of 4 triggers per day, representing a reduction of approximately 91.5%. Continuous effective power supply duration percentage refers to the ratio of the cumulative duration of continuous effective power supply and valid sampling data within the statistical period to the total duration of the statistical period. Under the same statistical conditions, the continuous effective power supply duration percentage was 86% for the control group and 98.2% for the experimental group. The experimental group reduced the amount of power supplied during the spark precursor stage, thus reducing the number of incidental events. Figure 7 The data continuity interruption caused by entering the hardware latch lock state is shown.

[0115] Exit dust concentration: Using the same outlet dust concentration detection device and the same statistical method, the average outlet dust concentration of the control group was 28.5 mg / Nm³. 3 The average dust concentration at the outlet of the experimental group was 16.2 mg / Nm³. 3 The experimental group had an average reduction of 12.3 mg / Nm compared to the control group. 3 .

[0116] The above operational results demonstrate that, under similar macroscopic operating conditions and identical hardware protection conditions, the method of this invention can identify spark precursors based on the finite-time maximum Lyapunov exponent input characteristics and cross-scale eigenvectors, and perform constrained power reduction regulation before the hardware overcurrent protection trips. Within the statistical period of this embodiment, the experimental group experienced fewer hardware blocking signal triggers than the control group, a higher proportion of continuous effective power supply time, and a lower average outlet dust concentration. These results indicate that the method of this invention can reduce power outages caused by frequent hardware blocking, while the asynchronous hardware protection still operates independently.

Claims

1. A method for controlling the operating conditions of a variable frequency power supply for an electrostatic precipitator based on deep learning, characterized in that, include: The evaluation network is based on macroscopic operational state data to determine the number of delayed sampling points, embedding dimension, and dynamic evolution threshold; The original high-frequency secondary voltage sampling signal and secondary current sampling signal acquired synchronously are processed to form a one-dimensional net voltage timing signal and secondary current correlation features; The phase space of the one-dimensional net voltage time-series signal is reconstructed by the number of delayed sampling points and the embedding dimension, forming a finite-time maximum Lyapunov exponential input feature, which, together with the secondary current-related feature, constitutes the microscopic electrical feature. The micro-feature gating coefficients are determined based on the input features and dynamic evolution thresholds, and the micro-electrical features and macro-operational status data are fused to form a cross-scale feature vector; The cross-scale feature vector, feedback control state value, and current power supply mode are input into the control network to obtain the micro-operating condition classification results and the initial set values ​​of the control parameters. The initial set values ​​of the control parameters are verified by deterministic boundary constraints and encoded to form control parameter commands.

2. The method for controlling the operating conditions of an electrostatic precipitator variable frequency power supply based on deep learning according to claim 1, characterized in that, Macro-level operating status data includes boiler load, dust collector inlet flue gas temperature and flue gas flow rate, and the evaluation network is a multilayer sensor. The candidate delayed sampling points, candidate embedding dimension, and candidate dynamic evolution threshold formed by the same macroscopic operating state update constitute an indivisible candidate dynamic parameter set; The stability condition is met when the number of candidate delayed sampling points and the candidate embedding dimension are the same in two consecutive macroscopic operating state update cycles, and the difference in the candidate dynamic evolution threshold does not exceed the preset threshold change tolerance. Once the candidate dynamic parameter set meets the stability condition, the preset historical data length requirement, and the preset control update cycle boundary condition, it will be activated as a whole in the next control update cycle. The three parameters will be used as the current effective number of delayed sampling points, the current effective embedding dimension, and the current effective dynamic evolution threshold, respectively.

3. The method for controlling the operating conditions of an electrostatic precipitator variable frequency power supply based on deep learning according to claim 1, characterized in that, Obtain each pre-divided continuous effective historical switching cycle, calculate the periodic average voltage of each switching cycle, subtract the corresponding periodic average voltage from each phase sample value to obtain the phase residual, and calculate the arithmetic mean of the phase residuals of the same phase sample point number to form a phase ripple template. The net voltage sample value is formed by subtracting the phase ripple template value of the corresponding phase from the current original high-frequency secondary voltage sample value. The net voltage sample value is then formed by digital low-pass filtering and preset integer multiple downsampling to form a one-dimensional net voltage timing signal.

4. The method for controlling the operating conditions of an electrostatic precipitator variable frequency power supply based on deep learning according to claim 3, characterized in that, After digital low-pass filtering and downsampling by the same factor, the secondary current sampling signal is aligned with the one-dimensional net voltage timing signal according to the sampling sequence number or timestamp. A time sliding window corresponding to the phase space reconstruction is established. The mean value of the secondary current is calculated within the time sliding window. The slope of the first-order linear fitting of the secondary current sample value and the sampling time is taken as the rate of change of the secondary current. The zero-delay Pearson linear correlation coefficient between the one-dimensional net voltage time sequence signal and the aligned secondary current sample signal is taken as the voltage-current correlation feature. The three together constitute the secondary current correlation feature.

5. The method for controlling the operating conditions of an electrostatic precipitator variable frequency power supply based on deep learning according to claim 2, characterized in that, Construct a historical voltage data buffer queue for storing net voltage signals, and select target net voltage sampling points within the same continuous valid data segment of the historical voltage data buffer queue; Taking the target net voltage sampling point as the end, select historical sampling points with the current effective embedding dimension reduced by one in the historical direction according to the current effective number of delayed sampling points, and form a multi-dimensional phase space state vector. The effective state vectors are arranged in chronological order to form a multidimensional phase space trajectory matrix.

6. The method for controlling the operating conditions of an electrostatic precipitator variable frequency power supply based on deep learning according to claim 5, characterized in that, The maximum Lyapunov exponent in finite time is calculated from the multidimensional phase space trajectory matrix. Based on the validity status of the calculation result, the input feature is formed by processing the result with valid values, the most recent valid value, or risk substitution values. When processing risk substitute values, the micro-feature gating coefficient is set to the preset maximum gating value; When risk substitution value processing is not performed, the same standardization processing is performed on the input features and the current effective dynamic evolution threshold, and the micro-feature gating coefficient is determined by the Sigmoid function based on the difference between the two. After the micro-feature gating coefficient is determined, the input features, secondary current related features and macro-operating state data are respectively standardized. The standardized micro-electrical features are multiplied by the micro-feature gating coefficient, and the standardized macro-operating state data are multiplied by one and subtracted from the micro-feature gating coefficient. The results are then concatenated to form a seven-dimensional cross-scale feature vector.

7. The method for controlling the operating conditions of an electrostatic precipitator variable frequency power supply based on deep learning according to claim 6, characterized in that, The control network is a multi-task feedforward neural network that includes a shared input layer, a shared hidden layer, a classification branch, and a regression branch. The inputs to the control network include a seven-dimensional cross-scale feature vector, three feedback control state values ​​corresponding to the fundamental voltage amplitude, pulse power supply frequency, and pulse duty cycle, respectively, and a three-dimensional one-hot encoding of the current power supply mode. The classification branch outputs the classification probabilities of normal corona, spark precursor, and reverse corona, which are then verified by confidence to form the micro-operating condition classification result. The initial settings for the fundamental voltage amplitude, pulse power supply frequency, and pulse duty cycle of the regression branch output.

8. The method for controlling the operating conditions of an electrostatic precipitator variable frequency power supply based on deep learning according to claim 7, characterized in that, The average power supply is determined by the power supply mapping table; When there are signs of sparking, it is forbidden to increase the fundamental voltage amplitude and average power supply. It is permissible to reduce the fundamental voltage amplitude, reduce the pulse duty cycle, or enter a low-power pulse operation state in pulse power supply mode. If the current power supply mode is not pulse power supply mode and the average power supply power of the preset low-power pulse control parameter group is not higher than the current average power supply power, the pulse power supply mode and the preset low-power pulse control parameter group are respectively regarded as candidate power supply mode and target control parameter. During back corona discharge, reduce the average power supply or pulse duty cycle; If the current power supply mode is not intermittent and the average power supply power of the preset safety control parameter corresponding to the intermittent power supply mode is lower than the current average power supply power, the intermittent power supply mode and the preset safety control parameter shall be used as the candidate power supply mode and the target control parameter, respectively. When the power supply mode switching is triggered by the spark precursor state or the back corona state, the micro-condition classification result continuously reaches the preset mode switching cycle number, and the current control system is not in the output blocking state. The preset equipment safety conditions corresponding to the candidate power supply mode are met, and the candidate power supply mode and the target control parameters meet the preset theoretical continuous effective power supply duration requirements.

9. The method for controlling the operating conditions of an electrostatic precipitator variable frequency power supply based on deep learning according to claim 8, characterized in that, Deterministic boundary constraint verification includes finite value judgment, physical quantity conversion judgment, single maximum change limit, absolute value limit, mode switching condition verification, and abnormal state interception. In the abnormal state interception, when the system detects the existence of any of the following trigger states: The following trigger states are triggered: invalid secondary current related characteristics, invalid maximum Lyapunov exponent in finite time, insufficient classification confidence, inconsistent control action direction, external macroscopic input distribution, spark precursor classification result, control network inference timeout, abnormal control parameters, deterministic boundary constraint verification timeout, and hardware blockade. Increasing the fundamental voltage amplitude is prohibited. In pulse power supply mode or intermittent power supply mode, the average power supply corresponding to the control parameter combination should not exceed the average power supply corresponding to the actual control parameters used in the previous control update cycle.

10. The method for controlling the operating conditions of an electrostatic precipitator variable frequency power supply based on deep learning according to claim 1, characterized in that, The method is applied to a hardware system including an inverter bridge, a fast comparator, and a hardware latch circuit. The method also includes hardware asynchronous protection. When the secondary current exceeds the hardware overcurrent protection threshold, the fast comparator directly outputs a hardware blocking signal to asynchronously cut off the gate drive of the inverter bridge switching device, and the hardware latch circuit maintains the hardware latch blocking state. Hardware asynchronous protection always runs independently and in parallel. Its protection priority is higher than control network inference, deterministic boundary constraint verification and software control parameter instructions. Deep learning algorithms cannot shield, delay or cover hardware blocking signals.

11. The method for controlling the operating conditions of an electrostatic precipitator variable frequency power supply based on deep learning according to claim 10, characterized in that, After the hardware latch block state is valid, the hardware block trigger state is recorded, the unexecuted power supply mode switching request is revoked, the use of the control network to generate new control parameter instructions is prohibited, the phase ripple template used to process the one-dimensional net voltage timing signal is marked as invalid, and a data continuity interruption flag is written into the historical data sequence used for phase space reconstruction. After the hardware latch is released, the phase ripple template is reconstructed according to the preset safety control parameters or mode switching safety control parameters corresponding to the current power supply mode, and continuous and valid historical data for phase space reconstruction is accumulated. Before the preset template reconstruction and recovery conditions are met, the control network is not used to generate new control parameter instructions.

Images